Research article.

Published in Open Differential Psychology 9th of October, 2014

Submitted 27th of August, 2014

Crime, income, educational attainment

and employment among immigrant
groups in Norway and Finland
Emil O. W. Kirkegaard1
Abstract
I present new predictive analyses for crime, income, educational attainment and employment among
immigrant groups in Norway and crime in Finland. Furthermore I show that the Norwegian data
contains a strong general socioeconomic factor (S) which is highly predictable from country-level
variables (National IQ .59, Islam prevalence -.71, international general socioeconomic factor .72, GDP
.55), and correlates highly (.78) with the analogous factor among immigrant groups in Denmark.
Analyses of the prediction vectors show very high correlations (generally > .9) between predictors
which means that the same variables are relatively well or weakly predicted no matter which predictor
is used.
Using the method of correlated vectors shows that it is the underlying S factor that drives the
associations between predictors and socioeconomic traits, not the remaining variance (all correlations
near unity).

Keywords: National IQ, intelligence, group differences, country of origin, Norway, Finland, Denmark,
immigration, crime, spatial transferability hypothesis, income, employment, educational attainment,
general socioeconomic factor, Islam, method of correlated vectors

1 Introduction
Recent studies show that criminality and other socioeconomic traits such as educational attainment
among immigrant groups is predictable from their country of origin[1, 2, 3, 4]. This study attempts
to replicate and generalize these findings.
The theoretical impetus for testing country-level predictors2 is the spatial transferability hypothesis.[5]
In brief, it proposes that 1) a countrys performance on a variety of metrics is due to some degree to
the psychological makeup of its inhabitants; 2) people retain their psychological attributes to some
degree when they migrate; and hence 3) the psychological attributes of groups determine to some
degree their relative performance on a variety of socioeconomic variables, such as crime, educational
attainments, income, and employment rate, in the countries that receive them.
For instance, when people from a poorer country move to a wealthier country, they will tend to
be relatively poor in that country as well. This is because part of the reason the country is poor to
begin with is that the people living there are low (or high) in the relevant behavioral traits. When
they move to a new country, they will generally still be low in the relevant behavioral traits, and
this will cause them to be relatively poor in that country as well. This is of course still allowing for
other causes (e.g. culture or religion people tend to bring with them) as well as improvements on
an absolute scale. Somalis living in Denmark are far richer than those who have stayed behind in
1
Department of Culture and Society, University of Aarhus. Corresponding author: emil@emilkirkegaard.dk
2
A note about terminology. predictor is used here to mean the same as independent variable. No causality is
implied, merely linguistic convenience.
Somalia, but they are nonetheless poorer than ethnic Danes, just as Somalia is poorer as a whole
than Denmark. Note also that the hypothesis does not specify why these traits tend to be preserved.
Both genetic and non-genetic models are possible.
The most obvious way of testing the spatial transferability hypothesis involves looking at immigrant
performance on a variety of measures grouped by country of origin, and then checking how predictable
this performance is from country-level variables such as national IQ and national prevalence of Islam.
In this study I explore a number of datasets in this fashion.

2 Dataset 1: Norway and Finland

Skardhamar et al (2014)[6] presented new crime data for immigrant groups by country and macro-
region of origin for Norway and Finland and compared the two countries. In doing so they adjusted
for both gender ratio and age structure in the populations. Their main findings are shown in Figures
1 and 2.

Figure 1: Violent crime in Norway and Finland by country of origin. From [6]

Figure 2: Property crime (larceny) in Norway and Finland by country of origin. From [6]
 These findings indicate that people from the same areas of origin are similarly disposed to criminal
behavior in Norway and Finland.

3 Dataset 2-4: Norway

I took a closer look at Statistisk Sentralbyras (SSB) 3 website for data that could be useful for testing
the spatial transferability hypothesis. To be useful, the data must concern a variable of considerable
social interest and contain information on immigrant group performance by country of origin with at
least a small sample (my threshold was 10) of countries.
I searched for landbakgrunn (country background) on the website4 , limited the results to
publicly available datasets (statistikbanken) and looked through all 124 results. I identified three
useful datasets:
1) Income after tax, measured as a percent of the national mean (income index).5 No information
about age is given. I included all available countries (N=23). Generally, SSB limits the available
countries to the ones with samples large enough to give reliable results.
2) Registered unemployed persons by sex aged 15-74, as a percent of the working population.6 As
before I included all available countries (N=120) and both sexes separately. There is no fine-grained
age information, so the data are not well-adjusted for age.
3) Tertiary educational attainment per capita for persons aged 16 and above in 2013.7 This
table was in absolute numbers, so I supplemented it with the population size by country of origin to
calculate a pseudo per capita value.8 The reason it is a pseudo per capita is that population sizes
were not available by age groups, so I had to use the entire age group, even though the educational
attainment data concerned only people aged 16 and above. This introduces error if the age structures
are different between groups. The data are also not broken down by gender, so there is possibly
gender ratio bias as well. To examine effects of including small samples, I used two different versions
of this variable. The first includes all groups with a population 200 (N=118). The second only
includes groups with 1000 to reduce sampling error (N=67).

4 Predictive analyses
I did all analyses with R.9
The primary question was whether crime was predictable from country-level variables as previously
found. To test this, I used the following predictors in a correlation analysis:

Prevalence of Islam in 2010 (as estimated by the Pew Research Center).[7]

Lynn and Vanhanens national IQs with changes based on the work of Jason Malloy. When a
value is changed, it is noted in the datafile.[8, 9]

Altinoks educational achievements.[10]

The World Banks GDP per capita (2013).[11]

Kirkegaards country-level general socioeconomic (S) factor scores.[12]

3
The official statistics bureau of Norway. http://www.ssb.no/
4
I could not find out how the SSB determines the origin country. Which country does a person who has a Swedish
and a Norwegian parent count as? Likewise there is no information about immigrant generation.
5
Tabell: 10489: Innvandreres inntekt etter skatt per forbruksenhet, etter landbakgrunn
6
Tabell: 07117: Registrerte arbeidsledige 15-74 ar, etter landbakgrunn og kjnn. Absolutte tall og i prosent av
arbeidsstyrken
7
Tabell: 09623: Innvandrere 16 ar og over, etter utdanningsniva og landbakgrunn. Absolutte tall
8
Tabell: 05184: Innvandrere, etter kjnn og landbakgrunn
9
R is a free, powerful, easy to use programming language designed for data mining and statistics. See http://www.r-
project.org/
 Table 1 shows the correlations of interest. Generally, all predictors do well when two conditions
are satisfied: 1) the sample of countries is large enough to have significant inter-country variation,
and 2) the sample of countries is not so large as to introduce significant sampling error in estimates.
The reason this introduces error is that the more countries covered in a variable means that the value
must be based on a smaller sample of persons from that country.

Table 1: Correlation matrix for country-level predictors and socioeconomic variables.

Findings of note include: Violent crime is easier to predict than property crime, just as in the
Danish dataset.[1] The poor predictive ability of Altinok with the crime and income variables seems
to be due to sampling error (Ns 13-14). The educational attainment variable which includes only
large samples (Tert. Ed. Att. Big) has higher correlations than the one with smaller samples too.
This is probably because the smaller ones introduce sampling error. Islam is a better predictor of
female unemployment than of male, which may be related to the role of women in Islam.
The intercorrelations between predictors is shown in Section 11.3 in the Appendix. IQ, Altinok,
logGDP and international S have high intercorrelations, with a minimum of .72 and a mean of .84.
Islam correlates weakly to moderately with the others (-.14 to -.43, mean -.29).

4.1 Predictor vector intercorrelations

Are some predictors just generally better at predicting than others, or is there specificity such that
while predictor A may be better at predicting outcome X, predictor B is better at predicting outcome
Y? An example of this would be that the prevalence of Islam predicts crime better than (national) IQ
at predicting crime, while IQ is better at predicting educational attainment.
To investigate this, I correlated the prediction vectors (rows in Table 1) for each predictor with
the vectors of each other predictor. Correlations at 1 indicate that predictor performance is general,
while correlations near 0 indicate specificity. Table 2 shows the results.

Var Altinok.cors Islam.cors GDP.cors S.cors

IQ.cors 0.83 -0.99 0.93 0.96
Altinok.cors -0.86 0.9 0.92
Islam.cors -0.92 -0.97
GDP.cors 0.97

Table 2: Correlation matrix of predictor vectors. N=9

Surprisingly, even though there are problems with small sample sizes of the predictive correlations
and the length of the vectors (N=9), the results strongly suggest that what is well-predicted by one
predictor is also well-predicted by other predictors, no matter which two were compared. The mean
abs. r=.92.
5 A general socioeconomic factor among immigrant groups
in Norway
Similarly to my previous study of immigrant groups in Denmark[3], I wanted to investigate the
possibility of a general socioeconomic factor at the group level (S factor).[12] To do this, I used all
the variables concerning Norway except for the educational attainment with smaller groups to avoid
duplicating variables.

5.1 Handling missing values

Factor analytic methods require that there are no missing values. The easiest and most common
way to deal with this is to limit the data to the subset with complete cases. This, however, produces
biased results if the data are not missing completely at random, which they rarely are. Furthermore,
it heavily reduces sample sizes. Lastly, it wastes non-redundant information and potentially resources
spent gathering it. If a case has values for 5 out of 6 chosen variables, removing the case wastes 5
pieces of useful information.[13, 14, 15, 16] Table 3 shows the distribution of missing values.

Number of missing values 0 1 2 3 4 6

Number of cases 15 3 8 41 61 141

Table 3: The distribution of missing values in the Norwegian dataset.

For the above reasons, I used four methods for handling missing cases: 1) complete cases only
(N=15), 2) imputing10 data in cases with 1 missing value (N=18), 3) imputing data in cases with 2
or fewer missing values (N=26), 4) imputing data in cases with 3 or fewer missing values (N=67).
Table 4 shows descriptive statistics for each dataset. The imputed datasets are similar to both the
full datasets and the complete cases although there were changes in both the skew and kurtosis.
10
I used the VIM package 4.00. The irmi() function imputes values.[17] I used the default settings. http://cran.r-
project.org/web/packages/VIM/index.html
 Var name Dataset n mean sd min max skew kurtosis
Violent crime Full 26 1.31 0.87 0.2 3.2 0.55 -0.83
Complete cases 15 1.41 0.99 0.2 3.2 0.39 -1.25
Impute 1 18 1.33 0.93 0.2 3.2 0.57 -0.94
Impute 2 26 1.28 0.81 0.2 3.2 0.7 -0.27
Impute 3 67 1.23 0.63 0.2 3.2 1.05 1.41
Larceny Full 26 0.77 0.56 0.1 2 0.56 -1.09
Complete cases 15 0.78 0.55 0.2 1.6 0.38 -1.74
Impute 1 18 0.72 0.53 0.1 1.6 0.55 -1.42
Impute 2 26 0.69 0.56 -0.29 1.96 0.62 -0.58
Impute 3 67 0.69 0.34 0.1 1.6 0.71 0.19
Tert. ed. att. Full 67 0.12 0.08 0.01 0.31 0.42 -0.91
Complete cases 15 0.1 0.07 0.01 0.23 0.39 -1.32
Impute 1 18 0.1 0.07 0.01 0.23 0.5 -1.12
Impute 2 26 0.09 0.07 0.01 0.24 0.75 -0.63
Impute 3 67 0.12 0.08 0.01 0.31 0.42 -0.91
Unemployment, men Full 120 7.05 4.18 1.38 22.08 1.26 1.69
Complete cases 15 7.4 5.36 2.68 22.08 1.38 1.18
Impute 1 18 7.31 4.89 2.68 22.08 1.57 2.16
Impute 2 26 6.88 4.3 2.66 22.08 1.8 3.72
Impute 3 67 6.81 4.11 1.66 22.08 1.39 2.18
Unemployment, women Full 120 7.5 4.97 1.32 31.82 1.92 5.11
Complete cases 15 8.9 6.2 1.98 22.42 0.83 -0.58
Impute 1 18 8.17 5.93 1.9 22.42 1.03 -0.04
Impute 2 26 7.4 5.25 1.56 22.42 1.3 1.14
Impute 3 67 7.41 5.39 1.32 31.82 1.97 5.23
Income Full 23 79.86 14.58 53.25 108.25 -0.01 -0.92
Complete cases 15 78.78 14.78 53.25 108.25 0.19 -0.78
Impute 1 18 82.4 16.3 53.25 112.29 0.11 -0.95
Impute 2 26 79.09 13.9 53.25 108.25 0.13 -0.75
Impute 3 67 81.71 13.86 31.96 108.25 -0.76 0.95

Table 4: Descriptive stats by dataset.

KMO tests show that all datasets are suitable to factor analysis, KMOs .68-.75. Note that the
method of imputation used is probabilistic, i.e. does not result in the same imputation every time.
Therefore, any researcher who replicates the analyses will find that the numbers deviate somewhat
from the shown results. The KMO values were always around these values in my tests.

5.2 Number of factors to extract

To find out how many factors to extract, I ran nScree() from the nFactors package.11 For each
dataset, all four tests within that function suggested to extract only 1 factor.

5.3 Strength of the general factor

Previous studies show that principal component analysis tends to overestimate factor loadings when
used on a small number of variables, but that other factor methods yield very similar results.[12, 18,
19, 20] I used minimum residuals (the default) to extract the first factor from each dataset.12
11
Version 2.3.3 http://cran.r-project.org/web/packages/nFactors/index.html
12
I used the fa() function from Psych package. http://cran.r-project.org/web/packages/psych/index.html Version:
1.4.8.11
 Revelle and Wilt[21] showed that one cannot solely rely on the size of the first factor in a normal
analysis as a measure of the strength of the general factor. They advocate five other methods, of
which I have used four here: 1) hierarchical omega and its asymptotic value, 2) the amount of
variance accounted for by the first factor in a Schmid-Leiman transformation, 3) the explained
common variance, and 4) the squared multiple correlation of regressing the first factor on the original
variables.13 Table 5 shows the comparison statistics including data from the reanalysis of the Danish
data presented in the next section.

Dataset Var% MR Var% MR SL Omega h. Omega h. a. ECV R2

NO Complete cases 0.68 0.65 0.87 0.91 0.78 0.98
NO Impute 1 0.66 0.62 0.86 0.9 0.74 0.96
NO Impute 2 0.64 0.60 0.85 0.89 0.75 0.95
NO Impute 3 0.63 0.59 0.82 0.87 0.73 0.99
DK complete cases 0.57 0.51 0.83 0.85 0.68 0.99
DK impute 4 0.55 0.51 0.86 0.88 0.73 0.99
International S factor 0.43 0.35 0.76 0.77 0.51 0.81
Cognitive data 0.33 0.74 0.79 0.57 0.78
Personality data 0.16 0.37 0.48 0.34 0.41

Table 5: Measures of general factor strength. The cognitive and personality data is from Revelle
and Wilt (2013)[21], the international S factor data is from Kirkegaard (2014)[12], and the Danish
comparison data is from a reanalysis of the datasets from Kirkegaard and Fuerst (2014)[3] presented
in the next section.

The data makes it clear that the S factors at the group-level among immigrants in Norway and
Denmark are very strong, even compared to the international S factor and the general factor of
cognitive ability (g) in 5 classic datasets. The imputation of data has little effect on the measures of
general factor strength.

6 Reanalysis of immigrant performance in Denmark

To better investigate the question of the strength of the S factor within another country, I repeated
the analyses discussed above on the dataset from Kirkegaard and Fuerst (2014)[3]. Since I used
the same methods on this dataset as I did on the Norwegian ones discussed above, I will keep the
description short.
I analyzed the data with the fa() and omega() functions just as before. Results are shown in
Table 5 above.
KMO is .73 in the complete cases dataset and .83 in the imputed dataset.

6.1 Handling missing values

There are a few missing values in the dataset. I used two methods to deal with this: 1) complete cases
(N=31), and 2) imputation via the VIM package for cases with 4 or fewer missing values (N=70).
Table 6 shows the distribution of missing values.

Number of missing values 0 1 2 3 4 23

Number of cases 31 9 23 6 1 1

Table 6: Distribution of missing values in the Danish dataset.

13
I used the omega() function from Psych package to extract the information.
6.2 Predictor vector intercorrelations
I repeated the analysis from Subsection 4.1 on the Danish data. The Danish data allows for a better
test of the general vs. specificity models. This is because the Danish data has more variables (25
instead of 9; the Appendix contains a list of the variables), and they include more countries (Ns
close to 70), and they are age controlled. Table 7 shows the correlations between prediction vectors.

Predictors Altinok Islam logGDP S.score

IQ 0.99 -0.96 0.98 0.99
Altinok -0.94 0.98 0.98
Islam -0.94 -0.96
logGDP 0.99

Table 7: Correlation matrix of predictor vectors in the Danish data. N=25

The correlations are even closer to unity than they are in the Norwegian data. The mean abs.
r=.97. This is probably because the error sources are smaller in these data.

7 Predictive analyses of S scores

I wanted to know how well S factor scores in Norway are predictable from the country-level predictor
variables. Table 8 shows the correlation matrix with the Danish scores as a comparison.

Variable DA S imp. NO S complete NO S imp. 1 NO S imp. 2 NO S imp. 3

National IQ 0.54 0.72 0.73 0.66 0.59
Altinok 0.55 0.26 0.31 0.2 0.6
Islam -0.71 -0.79 -0.79 -0.72 -0.71
log(GDP) 0.51 0.35 0.4 0.44 0.55
National S 0.54 0.7 0.71 0.63 0.72
DA S imp. 0.91 0.91 0.79 0.78
NOS complete 1 0.99 0.99
NO S imp. 1 0.99 0.99
S imp. 2 0.99

Table 8: Correlation matrix of predictor variables and S factor scores in Denmark (with imputed
values) and the four Norwegian datasets with varying amounts of imputation.

The results indicate that S factor scores are about equally predictable by predictor values in
the full Danish and Norwegian datasets. The size of the correlation decreases with the amount of
imputation and increasing sample size. This may be because the imputation introduces error or that
the correlations are artificially high due to sampling error. The only discrepancy is the predictive
power of the National S (.54 vs. .72). I dont have any good guess for why this is.

8 Method of correlated vectors

Arthur Jensen invented the method of correlated vectors (MCV) in 1983 to find out if g is responsible
for mean differences in measures of intelligence.[22, 23, 24] Today, the method is mostly used with g
(e.g. [25, 26, 27, 28]), and in context to mean differences, but it can be used for any latent variable
and in other contexts such as prediction of grade point averages.[29] I have previously used it on
international rankings data and found that the international S factor was driving the correlations
with predictors such as national IQs.[12]
 To apply the method, one correlates the indicator variables (IVs) loading on the latent variable
of interest with their correlations with the criteria variable. If the general factor is driving the
association and is positively correlated with the criteria variable, then the correlation between factor
loadings on it and the effect sizes of the predictor-criterion associations will be positive. However, if
the association is driven by the variance not attributable to the general factor, the correlation will
generally be negative. And it will generally be somewhere in between if the association is driven by a
mix of general and non-general factors.
Since the method relies on the IVs of the latent variable, it is susceptible to IV sampling error.
There are three main sources of error. 1) When the number of IVs is small, the correlation will be
unstable from sample to sample. 2) If the IVs are unrepresentative of the total population of IVs,
then the correlation can be biased either way. 3) If the variance in IVs loadings on the latent variable
is restricted, the correlations will be smaller.[30]
The NIV is quite small for the Norwegian data (6), but reasonable for the Danish data (25). The
standard deviation of loadings in the Norwegian and Danish datasets are .83 and .75, respectively, so
range restriction does not appear to be a problem. The IVs are reasonably representative of things
considered socioeconomically important, especially for the Danish dataset. Table 9 shows the MCV
correlations.

Predictor Norway (N=6) Denmark (N=25)

IQ 0.99 0.97
Altinok 0.95 0.95
Islam 0.99 0.99
logGDP 1 0.95
Int. S factor 0.99 0.96

Table 9: Results from method of correlated vectors applied to the S factor in Norway and Denmark.
Note that Islam prevalence has been reversed because it is negatively related to the S factor.

In every case, the result is close to unity. Strong correlations can result from nonlinearity in the
data, so I examined the scatterplots. However, they were all very linear.

9 Discussion
The simple predictive analyses give results similar to those found earlier. They serve as a successful
replication and generalization to two new countries.
The analyses of general factor strength show that the local S factors are generally very strong,
surpassing even the g factor and the international S factor. This is due in part to the grouped nature
of the data as group correlations tend to go towards 1 when there arent sampling errors or a
non-linear relationship.[31]
Surprisingly, the analysis of the predictor vectors show very high near unity intercorrelations.
The results were even stronger in the reanalyzed Danish data, which makes it probable that the
somewhat lower correlations in the Norwegian data are due to statistical artifacts. I interpret this
as showing that predictors are very general in their predictive ability. In accordance with Occams
Razor, causal theories of these correlations should be similarly general, not specific.
Applying the method of correlated vectors showed that the relationships between the socioeconomic
variables and the predictor variables were driven by the latent trait (S factor), not the remaining
variance (correlations near unity).
Generally the results strongly confirm the spatial transferability hypothesis.[5]
Limitations include the small sample sizes and the lack of adjustment for age and sex in some of
the variables. This probably introduces some bias in an unknown direction. Note however that the
Danish data is age-controlled, and yet the results are very similar to the Norwegian ones, showing
that bias due to age is unlikely to be a large source of error.
10 Supplementary material
All datasets and source code are available in the submission thread at the Open Differential Psychology
forum. Most of the data used in the study can be found in version 1.5 extra of the Worldwide
Megadataset.
The appendix contains a list of S scores by group in Norway and Denmark.

References
[1] Emil O. W. Kirkegaard. Criminality and fertility among Danish immigrant populations. Open
Differential Psychology, 2014.

[2] Emil O. W. Kirkegaard. Criminality among Norwegian immigrant populations. Open Differential
Psychology, 2014.

[3] Emil O. W. Kirkegaard and John Fuerst. Educational attainment, income and use of social
benefits among 71 immigrant groups in Denmark. Open Differential Psychology, 2014.

[4] Eugenio Proto and Andrew J Oswald. National Happiness and Genetic Distance: A Cautious
Exploration. IZA Discussion Papers, July 2014, 2014.

[5] John Fuerst and Emil O. W. Kirkegaard. Do National IQs Predict U.S. Immigrant Cognitive
Ability and Outcomes? An Analysis of the National Longitudinal Survey of Freshman. Open
Differential Psychology, 2014.

[6] Torbjrn Skardhamar, Mikko Aaltonen, and Martti Lehti. Immigrant crime in Norway and
Finland. Journal of Scandinavian Studies in Criminology and Crime Prevention, (ahead-of-print):
1--21, 2014.

[7] Pew Research. The Future of the Global Muslim Population, 2011. URL http://features.
pewforum.org/muslim-population/.

[8] Richard Lynn and Tatu Vanhanen. Intelligence: A unifying construct for the social sciences.
Ulster Institute for Social Research, 2012.

[9] Jason Malloy. Human Varieties Global IQ. URL http://humanvarieties.org/category/

hv-global-iq/.

[10] Nadir Altinok, Claude Diebolt, and Jean-Luc Demeulemeester. A new international database on
education quality: 1965--2010. Applied Economics, 46(11):1212--1247, 2014.

[11] The World Bank. Gdp per capita, 2014. URL http://data.worldbank.org/indicator/NY.
GDP.MKTP.CD.

[12] Emil O. W. Kirkegaard. The international general socioeconomic factor: Factor analyzing
international rankings. Open Differential Psychology, 2014.

[13] Yu-Sung Su, Masanao Yajima, Andrew E Gelman, and Jennifer Hill. Multiple imputation with
diagnostics (mi) in r: Opening windows into the black box. Journal of Statistical Software, 45
(2):1--31, 2011.

[14] Geert JMG van der Heijden, A Rogier T Donders, Theo Stijnen, and Karel GM Moons. Imputation
of missing values is superior to complete case analysis and the missing-indicator method in
multivariable diagnostic research: a clinical example. Journal of clinical epidemiology, 59(10):
1102--1109, 2006.
[15] A Rogier T Donders, Geert JMG van der Heijden, Theo Stijnen, and Karel GM Moons. Review:
a gentle introduction to imputation of missing values. Journal of clinical epidemiology, 59(10):
1087--1091, 2006.

[16] Kristel JM Janssen, A Rogier T Donders, Frank E Harrell Jr, Yvonne Vergouwe, Qingxia Chen,
Diederick E Grobbee, and Karel GM Moons. Missing covariate data in medical research: to
impute is better than to ignore. Journal of clinical epidemiology, 63(7):721--727, 2010.

[17] Matthias Templ and Peter Filzmoser. Visualization of missing values using the r-package vim.
Reserach report cs-2008-1, Department of Statistics and Probability Therory, Vienna University
of Technology, 2008.

[18] Randy G Floyd, Elizabeth I Shands, Fawziya A Rafael, Renee Bergeron, and Kevin S McGrew.
The dependability of general-factor loadings: The effects of factor-extraction methods, test
battery composition, test battery size, and their interactions. Intelligence, 37(5):453--465, 2009.

[19] Jason Major. The dependability of the general factor of intelligence: Why g is not a first principal
component. 2010.

[20] Jason T Major, Wendy Johnson, and Thomas J Bouchard Jr. The dependability of the general
factor of intelligence: Why small, single-factor models do not adequately represent g. Intelligence,
39(5):418--433, 2011.

[21] William Revelle and Joshua Wilt. The general factor of personality: A general critique. Journal
of research in personality, 47(5):493--504, 2013.

[22] Arthur R Jensen. Effects of inbreeding on mental-ability factors. Personality and Individual
Differences, 4(1):71--87, 1983.

[23] Arthur R Jensen and Li-Jen Weng. What is a good g? Intelligence, 18(3):231--258, 1994.

[24] Arthur Robert Jensen. The g factor: The science of mental ability. Praeger Westport, CT, 1998.

[25] Elijah L Armstrong, Michael A Woodley, and Richard Lynn. Cognitive abilities amongst the
sami population. Intelligence, 46:35--39, 2014.

[26] Hannah Peach, Jordan E Lyerly, and Charlie L Reeve. Replication of the jensen effect on
dysgenic fertility: An analysis using a large sample of american youth. Personality and Individual
Differences, 71:56--59, 2014.

[27] Jan te Nijenhuis, Birthe Jongeneel-Grimen, and Emil OW Kirkegaard. Are headstart gains on
the i g/i factor? a meta-analysis. Intelligence, 46:209--215, 2014.

[28] Michael Anthony Woodley, Jan te Nijenhuis, Olev Must, and Aasa Must. Controlling for
increased guessing enhances the independence of the flynn effect from i g/i: The return of
the brand effect. Intelligence, 43:27--34, 2014.

[29] Emil O. W. Kirkegaard. The personal jensen coefficient does not predict grades beyond its
association with g. Open Differential Psychology, 2014.

[30] John E Hunter and Frank L Schmidt. Methods of meta-analysis: Correcting error and bias in
research findings. Sage, 3rd edition, 2014.

[31] David Lubinski and Lloyd G Humphreys. Seeing the forest from the trees: When predicting
the behavior or status of groups, correlate means. Psychology, Public Policy, and Law, 2(2):363,
1996.
11 Appendix
11.1 S factor scores by country

Name ID S Factor in NO S factor in DK

Afghanistan AFG -1.09 -1.38
Argentina ARG 0.75
Australia AUS 1.03 1.13
Austria AUT 1.02 0.95
Burundi BDI -0.54
Belgium BEL 1.16 1.09
Bulgaria BGR 0.17 0.81
Bosnia and Herzegovina BIH 0.49 -0.91
Brazil BRA -0.34 0.46
Canada CAN 1.03 1.14
Switzerland CHE 1.13 1.12
Chile CHL 0.25 0.28
China CHN 0.61 0.63
Congo Rep. COG -1.07
Colombia COL 0.26
Czech Republic CZE 0.43 0.25
Germany DEU 1.04 0.85
Denmark DNK 1
Algeria DZA -1.52 -0.78
Egypt Arab Rep. EGY -0.24
Eritrea ERI -0.43
Spain ESP 0.52 0.79
Estonia EST 0.19 0.72
Ethiopia ETH -0.16 -0.59
Finland FIN 0.78 0.89
France FRA 0.97 1.1
United Kingdom GBR 1.14 0.85
Ghana GHA 0.03 0.16
Gambia The GMB -0.84
Greece GRC 0.61 0.61
Croatia HRV 0.54 -0.12
Hungary HUN 0.45 0.84
Indonesia IDN 0.33 0.13
India IND 0.63 0.53
Ireland IRL 0.88
Iran Islamic Rep. IRN -0.35 -0.69
Iraq IRQ -2.26 -1.65
Iceland ISL 0.76 0.55
Israel ISR -0.06
Italy ITA 0.86 0.77
Jordan JOR -1.19
Japan JPN 1.02
Kenya KEN -0.24 0.09
Kosovo KSV -0.43
Kuwait KWT -2.62
 Lebanon LBN -1.03 -2.03
Sri Lanka LKA -0.14 -0.75
Lithuania LTU -0.08 0.9
Latvia LVA 0.06 0.68
Morocco MAR -0.63 -1.03
Macedonia FYR MKD -0.19 -0.44
Myanmar MMR -0.27 -1.81
Nigeria NGA -0.53 0.34
Netherlands NLD 1.11 1.12
Norway NOR 0.84
Nepal NPL 0.75
Pakistan PAK -0.87 -0.68
Peru PER 0.1
Philippines PHL 0.58 0.36
Poland POL -0.02 0.46
Portugal PRT 0.54 0.63
West Bank and Gaza PSE -3.8
Romania ROU 0.31 0.7
Russian Federation RUS -0.44 0.45
Sudan SDN -1.52
Somalia SOM -3.06 -2.05
Serbia SRB 0.46 -1.93
USSR SUN 0.17
Slovak Republic SVK 0.42
Sweden SWE 1.03 0.77
Syrian Arab Republic SYR -1.62 -2
Thailand THA -0.03 -0.23
Tunisia TUN -0.82
Turkey TUR -0.52 -1.42
Tanzania TZA -0.25
Uganda UGA -0.34
Ukraine UKR 0.34 0.69
United States USA 0.97 1.26
Vietnam VNM -0.11 -0.58
Former Yugoslavia2 YU2 -1.61
Former Yugoslavia YUG -1.25
South Africa ZAF 0.73

11.2 List of variables in the Danish dataset

Variable name
All.crime.age.15.19
All.crime.age.20.29
Income.15.19
Income.20.29
Income.30.39
Income.40.49
Income.50.59
Income.60
Basic.school.15.19
 Basic.school.20.29
Basic.school.30.39
Basic.school.40.49
Basic.school.50.59
Basic.school.60plus
Long.tert.edu.20.29
Long.tert.edu.30.39
Long.tert.edu.40.49
Long.tert.edu.50.59
Long.tert.edu.60plus
Social.benefits.16.19
Social.benefits.20.29
Social.benefits.30.39
Social.benefits.40.49
Social.benefits.50.59
Social.benefits.60plus

11.3 Intercorrelations between predictors

Vars Altinok Islam logGDP International S

IQ 0.91 -0.27 0.72 0.86
Altinok -0.43 0.76 0.87
Islam -0.14 -0.33
logGDP 0.9

The low correlations between Islam and the others is not due to sampling fluctuation. Ns
from 116 to 198. The full correlation matrix can be found in the supplementary material correla-
tions Norway2014.xlsx.